HowNet Based Chinese Question Classification

HowNet Based Chinese Question Classification Dongfeng Cai1, Jingguang Sun1, Guiping Zhang1, Dexin Lv1, Yanju Dong1, Yan Song1, Chao Yu1 1 Natural Language Processing Laboratory, Shenyang Institute of Aeronautical Engineering P.O.box 118, No.52 North Huanghe Street, Shenyang, Liaoning, China, 110034 Email: cdf@ge-soft.com, sunjingguang@gmail.com , zgp@ge-soft.com , sam801025@163.com , dongyanju163@163.com , mattsure@gmail.com , yc089067@sina.com Abstract. Question classification is the first step that Question Answering System must dispose, the precision of question classification greatly affect the subsequent processes. In this paper, we present a new question classification method which uses HowNet as the semantic resource to extract features, and we use Maximum Entropy Model to implement the method. The results validate the effectiveness of this method: the classification precision of coarse classes and fine classes reaches 92.18% and 83.86% respectively. Keywords: Question Answering System; question classification; HowNet; Maximum Entropy Model; classification feature; 1 Introduction Question classification is the first step that Question Answering System must dispose, the precision of question classification greatly affect the subsequent processes. Research on English question classification started earlier. At first, the representative classification method is based on rules, and then Reference [2] used SVM (Support Vector Machine), Reference [3] used double-deck method and SNoW (Sparse Network of Winnow), Reference [4] used WordNet to process question classification, all got good results. But, it still don’t make an intensive study of Chinese question classification. Reference [5] used Bayes model based on syntactic structure parsing to process Chinese question classification, and the classification accuracy of coarse classes and fine classes reached 86.62% and 71.92% respectively. In this paper, we present a more efficient method of Chinese question classification which uses HowNet as the semantic resource to extract features, and use Maximum Entropy Model to implement the method. The experiment results show that the classification precision of coarse classes and fine classes reaches 92.18% and 83.86% respectively. 2 HowNet HowNet[6] is an on-line common-sense knowledge base unveiling inter-conceptual relations and inter- attribute relations of concepts by means of lexicons of the Chinese and their English equivalents. In HowNet, sememe is an unseparated semantic unit, and the HowNet altogether containing about 2200 sememes. In HowNet, we describe concepts by using KDML (Knowledge Database Mark-up Language), and use DEF expression to describe the definition of concept of KDML. DEF can describe the detailed semantic feature of words. For example: “生日：DEF={time|时间:TimeSect={day|日},{ComeToWorld|问世:time={~}}}” The first sememe of the word in HowNet is the sememe that first appears in the DEF. Here in the example, the first sememe of “生日” is “time|时间”. HowNet 2005 describes the semantic information of 81447 words and defines the conception with 157185 notes. It can commendably cover the words that 366 appear in present open domain QA system. So we choose HowNet as the semantic resources to analyze Chinese questions. 3 Question Classification Criterion There is still not a uniform Chinese question classification criterion. HIT_IRLab[5] has been engaged in Chinese question classification for a long time. This kind of classification can meet the demand of the practical classification requirement. Table 1 detailed presents …

Content truncated to 500 words